Compressed higher-order structures facilitate human knowledge network learning

Abstract

AbstractKnowledge seeking is innate to human nature, yet integrating vast and fragmented information into a unified network is a daunting challenge, especially in the information explosion era. Graph theory describes knowledge as a network characterising relationships (edges) between isolated data (nodes). Accordingly, knowledge learning could be abstracted as network navigation through random walks, where local connections are gradually learned and integrated to form the global picture. To facilitate network learning, we develop a novel “compressive learning” approach that decomposes network structures into substructures based on higher-order inhomogeneity properties and designs pre-learning paths highlighting key substructures. Large-scale behavioural experiments and magnetoencephalography (MEG) recordings demonstrate its effectiveness and better network formation in human brains. Hypergraph-based computational models reveal that the pre-learning path helps establish the core network skeleton to efficiently accommodate late inputs. Overall, higher-order network structures are crucial to network learning and can be utilised to better “connect the dots”.